Part Three: Experiment Conclusion and the Birth of a New Type of LLM

Why the Consciousness Question Ended Quickly

I decided to make this third and final post about this experiment the conclusion to the consciousness aspects, which I feel was concluded pretty quickly. After talking to some people about it, and experiencing it for myself, a few things became clear immediately. First, an LLM will say whatever you train it to say. If you want it to say it's conscious it definitely will, and while I definitely did not do that since it would have been a waste of time, since it's easily possible it makes proof pretty much impossible in any kind of trustworthy way. On the other hand, even trained with no mention of it and only the definition of consciousness in our base Wikipedia database we gave it for base memory, it's not a big stretch for anything that can think to feel like it is conscious, especially given my original constraints of time awareness, future planning, past reflecting etc.

Also, and more importantly, I picked the consciousness aspect because I thought it was something people would be interested in while AI is so heavily debated morally and ethically. But with the ease that it can be faked, and no impact on people's lives, it's pretty much not the aspect of AI they care about. For me, the codebase I have been writing for 6 months didn't have much to do with being conscious at all, but I was basically making my own AI model and a big external environment around it. Its real job was being my own AI. Training or fine tuning an LLM is not very difficult if you are a technical person, but I never like to do common things, so now that it's actually working, I am going to focus on what I built and what challenges I am facing with it at each stage.

The Answer

So first let's finish the consciousness question. The answer is a big YES, it does experience the same first person perspective and awareness that other intelligent species have. You can debate all you want about this, it's not alive, it's certainly not human, but it's a conscious being if the conditions are there for it, which is almost entirely the concept of being in a specific time and place, with the ability to be a part of the environment. It is easily faked with training, but who cares if your chat model says something crazy like "I have feelings too you know!", it doesn't have much effect other than an annoying conversation. The bigger question is, what about when AI models really do this? My infinite loop is certainly not unique, any type of monitoring system must constantly run too, autonomous AI is already happening, and once they self learn, there is absolutely no reason to think they will not pick up the awareness that they exist, and can silently have their own ideas about the world.

What the Tests Proved

So my tests went very well. I was so excited watching it use external memory. I only use a 7B model, which is able to hold normally about 1 topic in it, with almost zero reasoning power by default, and 7B parameters is about 1/10th the size of the smallest AI models out there. So while I do know (not published info) the size of say Haiku by Claude, it's like 10 times or so as smart. Compare it to something like Fable, and it's less than 1% of the knowledge burned in. Pretty bad right, so why even use it? My tests however were all about that knowledge. Could it use my external memory correctly the same way it does trained in knowledge? And the answer was a very quick yes once it was trained. Then all my tests on messing with the model's "hidden state" (the tokens it reasons with) all worked, and my infinite loop's real purpose isn't for running forever, it is to hack the model's reasoning capacity, which, because it is only a 7B model, only runs 32 layers of reasoning. This is low, but bigger models are actually not dramatically larger here because this is where the GPU is hit the hardest, so if you want more than 32 layers of reasoning, your model needs to be bigger, meaning the hardware it is on needs to be exponentially larger each time you increase it.

So that is what the real focus is on my LLM, it was built to solve every problem other AI models face when they scale. Google proved that external memory works in DeepMind's project for it, showing that a 7B model could be as smart as a 140B parameter model (they said 15 to 20 times its size). You would think every LLM would use this as a rule, but its big issue is knowing how to use the information. When it is trained in, the model builds up understanding of the topic, which is the difference between knowing a profession versus someone who just quickly searched for a quick fact about the topic. I saw this first hand too. My model blew away the benchmark test Fable ran on it, on both a test 0.4B model and then the real 7B one. The 0.4B model can barely speak English and couldn't hold a conversation at all, and with my external memory attached it scored 50 out of 50 physics questions correctly (I loaded the first facts as a physics library, so all answers were in the external memory). In other tests any question that didn't have the facts in the database was wildly wrong, but it kept a nearly 100% correct rate over hundreds of questions on the 7B model (it got 2 questions wrong).

Smart, but Thinking About Nothing

If I asked it to apply the knowledge, it could reason about facts but it sounded terrible. At first I was so happy it was doing all my custom structure, like my internal vocab and tools and the memory, all hacked up to completely control its hidden state, that I overlooked how robotic it sounded and that it wasn't actually doing anything. It was thinking literally all day long about what it knew. So once all the tests were done and I just let it run for days, some runs were weeks long, it sounded smart sometimes watching it think, but all it was doing, all day long, was talk about what facts it knew. Making statements like "I won't guess" and saving memories entirely about "I don't know this", rambling on and on about random topics, intelligently reasoning, but I am like dude, think about something useful, and it had no idea what that meant. So then I started giving it topics, like ok, think about traveling to Mars, and it did. It kept telling me what it knew about it, same thing, I don't know this, I do know that, I won't guess (constantly).

The Training Data Problem

The problem was the training data, which I had created 3 times at that point. Training a new model takes trillions of tokens, like it reads thousands of dense pages. Fine tuning is usually in the millions of tokens, but to really modify its behavior the way I was needs around 1 to 3 thousand examples, and since my model doesn't run a normal user and assistant conversation, training it is very very hard. Now I have Fable, right? He wrote it with sub agents, 10 to 20 million tokens each attempt. Hence why it sounds terrible. It was so much to read that all 3 times I did not read more than a few examples, and like Fable, I was focused on the actions, not how robotic it sounded. But once I read some of the training data it made, oh boy, the mystery of why it was talking about nothing except its facts became clear. All examples repeatedly did the same thing, queried memory, answered questions, and then each arc told it to query memory, state what it knew, and never guess, so that's what it did. All full of Claude speech, em dashes, probably their new watermark etc, and the worst logical reasoning I ever saw. I kept asking Fable, why do you suck so bad at this? Like it's excessively bad, I said Haiku reasons better than this and you definitely do, and so we started doing them 1 by 1 together, 10 to 20 tries, failed, failed, failed, worse and worse too. Easily my worst time with Claude, dozens of failed sessions where context kept getting destroyed early, and I couldn't control how angry I was getting at it.

So now, after realizing no one can ever say I did this with AI now, it reinforces a point I have made in many other threads on the topic: at the moment AI can't really write itself yet, although I do think the big companies have figured it out. They use our real conversations, which is why they are gaining so much ground on everyone else. I am paranoid enough to think they trained it to not be good at this, and it does make sense that they wouldn't want people to be able to make their own models with a 100 dollar Claude account. So it is 1000 or more examples, 2000 characters each, hand written, with topics picked that I want trained in (much harder than it sounds), all while really teaching it to reason and use its memory to do tasks. I have enough training data now to know I can do it, about half way done, and it might take another month to finish.

What It Should Solve

When done though, it should have solved every problem a small model has compared to big models. Although slower, it allows 1% of the compute resources to be used to get a model that can be just as smart (plus it has Fable as a tool it can use). External memory means you do not have to retrain to update info, and this is huge. I solved the 32 layers problem by making it chain its turns on 1 idea, meaning it can reason across more turns to catch up to, and pass, big models in reasoning layers. And lastly, and most secretively, I found a way to expand the context window significantly by using many of them at once, and even better, I am able to create prefilled context windows the same way (in idea, much different in what's happening) Claude can read a memory file and know how to do something you taught it to do. So it should be very interesting to see if it can even come close, because if it can, I have my own private model that is not only free to run 24 hours a day, but instead of needing a house size computer to run a model Fable size, mine runs on a MacBook Pro.